Advertisement

Think Locally, Act Globally: Federated Learning with Local and Global Representations

阅读量:

Think Locally, Act Globally: Federated Learning with Local and Global Representations

  • NeurIPS 2019 workshop

1. Method

1.1 LG-FedAvg

  • 通过各节点的本地数据X_k, Y_k训练local representation(其中Y_k可以是label,也可以是X_k本身,还可以是自定义的label) H_k,auxiliary网络a_k,以及本地的global网络g_k
  • server端通过加权的方式(n_k/n),汇总各节点的g_k,得到global aggregation modelg

1.2 Fair Reperesentation Learning

  • a_k平行结构,训练对抗网络(看的不是很懂。。。)

2. 优势

  • 只通信部分网络g_k的参数,大大减少了通信量

全部评论 (0)

还没有任何评论哟~